What is a neural network?


A phd student at Researchgate asked a very interesting question. “What is a neural network?”. Before I answer the question directly, it is important to know more about the phd student and her background knowledge, because it makes a difference if a newbie asks such a question or an expert. As far I can see from the researchgate profile, the student comes with a background in economics and is interested in decision making in complex situations in market environments. So I have to adapt the answer a bit into that direction.

Answer: Neural networks are used as a replacement for individuals. Instead letting real humans play a game, the aim is to play the game automatically. Neural networks are implemented on computers. Their main advantage compared to programming languages like Java is, that neural networks are error-tolerant. That means, they can decide under noise. Most projects with neural networks which were done in the past ended in a project failure. That means, a typical neural network project is not able to replace human players. Most researcher believe, that they better understand how to program neural networks than 20 years ago and they want to do more projects which might be a success.


Answering questions on Researchgate

In advance to bring the different science-communites together, I want to answer today the first question from the Researchgate network. In a read-only access together with a remote comment it is possible to give the answer here in my blog. So I searched for a recent question which was posted 7 days ago: [1]

It was asked by a member of the National science foundation, and 5 users has answered it so far. According to the Researchgate guidelines, all users are members of a serious universities. But perhaps that is not very important and we can answer the question on a topic-based subject? Yes, it is possible and here is my answer.

State of the art image recognition uses domain-ontologies which is combined with traditional techniques likes Convolution neural networks. An ontology is equal to a taxonomy and stores knowledge about a subject. In the area of image recognition, that is knowledge about shape of the objects and their positions. For example, if the aim is to recognize cat-images, than the ontology has a templates for different cats. Building such ontologies from scratch is equal to program a computer-based-learning scenario in which a human-operator learns to label images manual. Different cases are presented and the human user must answer what he sees on the picture. For evaluating his answer, the newly created ontology is used.. Such ontologies can be used also for automatic recognition. They help to reduce the error-rate of traditional neural networks. Further literature about the topic can be found with a Google Scholar request “Image recognition ontology”.

How to get an account on Researchgate?

A small comparison between Facebook, Google Plus and Academia.edu has shown, that none of them is a valuable Social network. The quality in Facebook is too low (it consists mostly of funny clips), Google Plus is more serious but there is no traffic. Academia.edu is a well known Academic social network, but most papers are not from nature-science. So the remaining candidate is Researchgate. They are focussed on computer science and medicine, they have a huge amount of traffic and they have discussions with the option to upvote a researcher. The only problem with Researchgate is, that not everybody can be a member.

I have tested out the startpage. They have a button for “create account”, but then it is a bit difficult. If the user selects “I’m not a researcher” than he can’t create an account. If the user selects “I’m a researcher” in the hope that he pretends so, than researchgate wants to know the exact e-mail address of the company somebody is working. So the intention is to only let real researcher in the network and hold all trolls and fake-scientists out.

What does that mean? It results into the same filter like on Arxiv.org. If somebody has an e-mail address in a university or in a research company, he has no problem for getting an account. And most of them have such an account. It is hard to find a well known author how have written important papers about robotics, who is not member of researchgate.

So what are the options? The first alternative to Researchgate is Academia.edu which is open for everybody. It is possible to login with a normal Google account, and as far i can see, normal people have done so in the past. The second option is to use the remote comment function of a weblog to became part of researchgate without being an official member. Reading in the discussions is possible. So the url to an thread can be copy & pasted in the own WordPress blog and the answer is posted there.

I know, both alternatives are not very pleasant. But it is better than nothing. The discussion board of Researchgate starts with “https://www.researchgate.net/post/” so a simple Google search filter is all the needed information. According to the following thread https://www.researchgate.net/post/How_to_create_own_profile_and_account_in_Research_Gate_for_my_co-author_Roshchina_VD it is possible to send an invitation to a person. And here https://www.researchgate.net/post/How_many_of_you_have_an_account_on_Researchgate_and_Academiaedu_in_the_same_time is a discussion about the pros and cons of Researchgate vs. Academia.edu

Like i mentioned above the main difference is, that Researchgate has discussions, and has more traffic. And here https://www.researchgate.net/post/Are_there_fake_identities_at_ResearchGate is a discussion about fake-profiles and how to spot them. As far as i can see, the real Researchgate members have the possibility to flag a fake-profile and ban it from the network. Another option is to use ORCID for check the credentials.

The reason why Researchgate is so focussed on real names and real scientists has to do with a lack of a public peer-review system. That means, apart from Openwetware, there is no public wiki which has the aim to verify done research, so asking for the authors name is the only option the community has today. The idea is perhaps, that if only people who have a degree in science are allowed to publish in science, this will prevent spam.

And i would agree to this hypothesis. Without a peer-review wiki it is not possible to spot Fake-papers. To open up Researchgate for everyone, first a peer-review wiki has to be build from scratch. But, currently they are no projects which aim to do so.

A bit desinformation is also given in the Researchgate forum. Here https://www.researchgate.net/post/Is_it_essential_to_have_an_institutional_email_id_to_join_research_gate some guy has the opinion, that it is possible to register on Researchgate with a normal Gmail adress. But, If we are looking in the profile of the guy who has said so, he is:

• working at a university
• has a degree in PhD, B.Sc, M.Phil
• and has a long publishing list

In contrast, I have found no example, that somebody outside of the university sector has an researchgate account. On Academia.edu it is possible, on Researchgate not.

Einführung in ResearchGate

Neben dem Paltzhirsch Academia.edu gibt es noch weitere Plattform welche sich als Facebook für Wissenschaftler bezeichnet: ResearchGate. Da dieses Portal zeitlich früher am Start war ist es klug sich das ganze näher anzuschauen. Interessant ist die Motivation die hinter dem Projekt steckt. Keineswegs ging es darum, Exzellenz zu versammeln, sondern ResearchGate wurde mit der Absicht ins Leben gerufen “negative Forschungsergebnisse” (O-Ton des CEO) zu sammeln. Damit sind Paper gemeint über gescheiterte Projekte, umgeschüttete Glaskolben, vermurkste Befragungen und Experimente wo man dachte überlichtgeschwindigkeit erreicht zu haben, was dann aber doch nicht so war. Kurz gesagt, all jene Forschung die traditionell die Qualitätskriterien nicht erfüllt. Es verwundert nicht, dass ResearchGate in Germany den Hauptsitz hat, weil Deutschland bekanntlich dasjenige Land ist, was die meiste Erfahrung mit gescheiterter Forschung hat. Aus Deutschland kommt nicht nur das Transrapidprojekt, sondern hier wurde auch OS/2 eingesetzt als bereits klar war, dass das Betriebssystem ein Flop ist. Aber auch der Flughafen BER (bis heute nicht fertig und eine riesige Bauruine) ist made in Germany. Das heißt, in Deutschland gibt es sehr viel Bedarf nach einer Plattform wo man gescheiterte Projekte hochladen kann.

Die spannende Frage ist jetzt: was kann man tun um ResearchGate selbst zum Scheitern zu bringen. Eine Methode wäre es, wenn auf der Plattform copyright-geschütztes Material gelangt, wodurch Elsevier dann einen guten Grund hat, die Firma auf Schadensersatz zu verklagen wegen entgangener Gewinne. Eine weitere Möglichkeit ist, dass ResearchGate keine Sponsoren findet, sich aber nicht traut Werbung zu schalten und am Ende auf den Personalkosten sitzen bleibt und Konkurs anmelden muss. Aber diw wohl beste Möglichkeit wäre, wenn die User abwandern zu Academia.edu weil es dort viel spannender ist und ResarchGate nicht mehr hip genug ist. Oder, was auch gut denkbar wäre — besonders in Deutschland wo man nicht sehr aufgeschlossen Startups gegenüber ist, dass ResearchGate Gegenwind aus der Politik bekommt.